mode choice pivot desifn

[Viktoriya]

Dear Ngene-Team & Forum Participants,

I am currently trying to generate an efficient design for a mode choice study based on priors estimated in a pilot study. It is a pivot design including RP and SP parts. Since I do this for a first time, I face some difficulties.
In the RP part of the pilot study, the participants have to report which mode of transportation they usually use for their commuting trip (Options: pedestrian, bike, car, public transport), trip length (in Kilometers) and trip Duration (in Minutes).
In the SP part of the study, in order to present realistic situations to the participants, they become choice sets with the same 4 options (pedestrian, bike, car, public transport) where TIME for each alternative is estimated by 1. reduce or increase of the time for the used alternative based on the reported time (e.g. person 1 use a car for commuting and drive 20 Minutes => TIME for car in the choice sets is = -30%/-10%/+20% of the reported time; 2. Reduce or increase of the time for all other alternatives based on the reported kilometers (e.g. if person 1 drives 20 km to work, than it will take the following time with public transportation: -30%/-10%/+20% of an assumed average speed for public transportation * reported kilometers).

Since the participants can choose a different mode of transportation, I have some trouble to define the reference alternative for the pivot design.
My Ngene code looks like this so far:

; alts = PED, BIKE, PT, CAR
; rows = 40
; block = 5
; eff = (mnl,d)
; model:
U(PED) = b[-0.562] + b2[-0.363]*TIME_PED.piv[-30%,-10%,+20%] /
U(BIKE) = b[-0.910] + b3[-0.0481]*TIME_BIKE.piv[-30%,-10%,+20%] + b4[-0.0981]*ACC[2,5,10] /
U(PT) = b5[0.00428]*TIME_PT.piv[-30%,-10%,+20%] + b4*ACC + B6[-0.114]*WAIT[2,5,10] + b7[-0.614]*COST_PT.piv [-30%, -10%,+20%] /
U(CAR) = b[-0.601] + b8*TIME_CAR.piv[-30%,-10%,+20%] + b4*ACC+ b11[-0.408]*COST_CAR.piv[-30%,-10%,+20%]
$

I have the following questions:

- Is it possible not to have a reference alternative in the pivot design and if not, should I define the model as if it wasn´t a pivot design or how can I define the reference alternative in this case?
- Are the utility functions still correct with alternative specific parameters (even if they are not that good at this point) since the modes of transportation are actually define in sawtooth as attributes (we use a so called CBC advanced design module where mode of transportation is a primary attribute, the time and costs are conditional attributes and the waiting and access time are common attributes).

Thank you very much in advance for your help!

victoria

[Michiel Bliemer]

1. Currently we only support pivot designs where the reference is a separate alternative. So this may not suit your needs and therefore you may be better off to make a library of designs which we did in a a mode choice study for the UK. I would suggest you create a design for {short,medium,long} distance with used modes {ped,bike,pt,car}, so in total you create 12 designs with fixed levels. Then you can convert these fixed levels into percentages that you can apply to the actually reported levels, although you could also choose to use the fixed levels. You probably also do not want to show all modes for each trip, for example, for long distance trips, you will not show 'ped', so it makes sense creating different designs.

2. Yes your utility functions look fine.

Michiel

[Viktoriya]

Thank you for the answer!

[Viktoriya]

Dear Ngene Team,

we created as proposed different designs for the short, medium and long trips, however the Company that implemente the designs for us somehow cannot make individual designs and Need one design for all participants. So we face different Problems: 1. the trip lenght in the Population follows rather a exponential and not normal Distribution and 2. we suggest that it will be maybe wrong to choose an averge trip for estimating efficient design 3. it is maybe also wrong to write a Kind of extreme values for the attributes (e.g. Attribute = Time for Bike instead of Attribute Levels = -30%,-10%,+20% of for example an average trip of 7 km with 15km/h = [19.6,25.2,33.6] => Attribute Levels = extreme and average possible values of the time, e,g,[7,25,60]

What we tried in Ngene is instead of Changing the Attribute Levels to Change the Priors in order to consider also the variance of the trip lenght Distribution:

e.g. U(BIKE) = b3[(n,-0.803,0.516)] + b4[(n,-0.644,0.622)]*TIME_BIKE[0.7,0.9,1.2]+ b5[(n,-0.134,0.0256)]*ACC[2,5]/

where the mean of b4 mean = Beta for Time_Bike * Average Duration per trip and Beta Standard Deviation = square root (Var(Beta,TIME_BIKE)), where Var(Beta,TIME_BIKE) = Var(Beta)*Var(TIME_BIKE)+Var(Beta)*(E(TIME_BIKE))^2 +Var(TIME_BIKE)*(E(Beta))^2

TImE_Bike is [0.70.9,1.2] since it is a Pivot design where the time is -30%,-20%,+20%

Unfortunately since the StD is quite high or the choice probabilities are maybe too extreme Ngene is not able to find an efficient design.

I would be very very thankfull for a sugestion how or whether can we solve this Problem in Ngene?

Thanks in advanced!

Victoria

[Viktoriya]

P.S.

Another solution to implement it in Ngene that I tought about is to create one desgin for all trips but by adding constrains to consider short vs. medium/long trips.
E.g. If I have for BIKE => Time (short trips) = [5,7,9] and Time(medium/Long trips) =[42,54,72],
CAR => TIME (short trips) = [2,3,8] and Time(medium/Long trips) =[15,20,27]

than I tried to write constrains where
;cond:
if(BIKE.TIME_BIKE=[5,7,9],CAR.TIME_CAR=[2,3,8]),
if(BIKE.TIME_BIKE=[42,54,72],CAR.TIME_CAR=[15,20,27])

...
; model
U(BIKE) = ... b2*TIME_BIKE[5,7,8,42,54,72]/
U(CAR)= .... b2*TIME_CAR[2,3,6,15,20,27] ...

Would that be a better solution to create Efficiency design which somehow fits for both short and medium/Long trips?

Should I have rather small number of blocks (e.g. 3) or greater number of blocks (e.g.6)?

Thanks in advance for an advise!

V.

[Michiel Bliemer]

(Apologies for the late reply, I am currently on study leave)

I am not sure I understand how changing the priors will help you achieve what you want. You seem to propose Bayesian priors to overcome the problem, but I think you really need to find a solution in the attribute levels.

I think there are two solutions you can choose:

1. Create a single pivot design for an average population and apply to all respondents (possibly with some post processing in which you round off certain values or put limits on them)

2. Create multiple fixed designs. I do not understand why a company cannot have 3 versions of the survey, one for short distance, one for medium distance, and one for long distance. This would require to create 3 different designs in Ngene separately, and depending on the distance class, the respondent sees one of these 3 versions. This is the most common way of doing this. For the UK we assisted in creating a whole library of designs (thousands) in which we simply picked the appropriate design based on the characteristics of the trip and person.